Abstract: Aiming at the problem of  the small sample size of high-resolution remote sensing images and  traditional optimized support vector machine (SVM) algorithms easily falling into local optima and slow optimization speed, we  proposed an algorithm based on deep transfer learning and lion swarm optimization SVM (LSO-SVM) to classify remote sensing image scene. Firstly, after enhancing the image through adaptive contrast,  color aggregation vectors were used to extract image color features. Secondly, three kinds of pretrained networks were used to extract the transfer learning depth features of images. Finally, the manually extracted image features and the features obtained using three pretrained networks were fused by using a series of feature fusion methods, and inputted them into LSO-SVM for image scene classification. The results show that the algorithm solves the problems of difficulty in deep learning training in small sample situations and the tendency of traditional optimized SVM algorithms to fall into local optima and slow search speed. Under 80% training conditions, the classification accuracy of UCM Land-Use and RSSCN7 datasets reaches 99.52% and 98.57%, respectively.

Key words: remote sensing image, image classification, transfer learning, lion swarm optimization algorithm, color coherence vector